Multimodal perception researcher: LiDAR hallucination benchmarks, VLM tool use, and geometry-distilled VLA models
3
papers here
3
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Not yet scored: no publication counts, venue list, or yearly totals are available. This axis would measure total research productivity — papers per year and venue mix — once a fuller bibliographic record is linked.
Impact—
Not yet scored: citation counts, h-index, and normalized influence are unavailable in this snapshot. This axis would quantify how widely the reviewed papers are cited and built upon relative to field baselines.
Focus0/100
The focus score is 0/100, and the raw field distribution explains why: the three reviewed papers split evenly, one third each in robotics/robot learning (vision-language-action models), computer vision (LiDAR scene understanding for autonomous driving), and machine learning (multimodal vision-language agents). This is deliberate breadth — benchmarking, credit assignment, and distillation — rather than concentration; a longer record would show whether the spread persists.
Collaboration62/100
The collaboration score is 62/100, built on 22 distinct coauthors across the three reviewed papers. The recurring network includes Di Wen (listed twice among frequent collaborators), Zhiyuan Gao, Yanxiang Zhan, Mohammad Khoshnazar, Runyi Yang, Ruiping Liu, and Michael Beetz. Peng also shifts authorship roles — first, middle, and last positions — though no cataloged paper lists them as corresponding author.
Activity—
Not yet scored: no year-by-year publication counts or submission cadence are available. This axis would track publishing rhythm — papers per year and gaps between releases — across a full career record.
Reproducibility70/100
Reproducibility scores a mean of 70/100 across three papers. Evolve rates the LiDAR-Hallu benchmark as high readiness — a rigorous protocol, though only one model family was tested — while the VLM tool-call credit study and FOCAL-VLA are medium, the latter with thin statistics. Benchmark-first work with explicit protocols anchors this above-average score.
Research summary
Kunyu Peng (Tongji University) works on multimodal perception, with Evolve-reviewed output split evenly across three areas: LiDAR scene understanding for autonomous driving, vision-language-action models for robot learning, and multimodal vision-language agents — each one third of the cataloged papers. All three are 2026 arXiv preprints, and Peng occupies different authorship slots in each: first author of a counterfactual credit-assignment study that pays only for visual tool calls a VLM actually needed and used, lifting matched-arm accuracy by 6–9 points; last author of the LiDAR-Hallu benchmark exposing answer-bias and relation-blindness in two B4DL LiDAR-LLMs; and a middle-position co-author on FOCAL-VLA, which distills VGGT geometry and Track4World motion into π0.5 for subtask-guided manipulation. A recurring theme is testing whether multimodal models truly perceive rather than exploit shortcuts — through hallucination probing, counterfactual evaluation, and geometric grounding. Evolve rates reproducibility at a mean of 70/100 across the three papers (one high, two medium readiness) and collaboration at 62/100 with 22 distinct coauthors. Output, impact, and activity metrics are not yet available; this profile is name-resolved from public bibliographic data.
Collaboration pattern
Across the three cataloged papers, Peng worked with 22 distinct coauthors, yielding a collaboration score of 62/100. The recurring cluster includes Di Wen (appearing twice in the frequent-coauthor list), Zhiyuan Gao, Yanxiang Zhan, Mohammad Khoshnazar, Runyi Yang, Ruiping Liu, Yufan Chen, as well as Jeroen Schäfer and Murat Akkoyun. Authorship roles rotate across first, middle, and last positions; none of the three papers lists Peng as corresponding author.
Research keywords
Highlights
LiDAR-Hallu: a benchmark exposing hallucination in LiDAR language models
Counterfactual credit for VLM visual tool calls
FOCAL-VLA: geometry distillation for vision-language-action models
Timeline
- 2026Last author on the LiDAR-Hallu study 'Do LiDAR Language Models Really Understand Spatio-temporal Relationships?' (arXiv), probing two B4DL LiDAR-LLMs.
- 2026First author on 'When Should a VLM Look?', introducing counterfactual need-and-use credit for visual tool calls with 6–9 point matched-arm accuracy gains.
- 2026Middle-position co-author on FOCAL-VLA, adding subtask-guided geometry distillation and implicit world modeling to vision-language-action models.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric dimension scores are computed by Evolve from public bibliometric data; no citation counts, h-index, or venue-tier metrics were provided.
- Scope is limited to three Evolve-reviewed 2026 arXiv preprints; earlier and peer-reviewed publications are outside this snapshot.
- Profile is name-resolved from public bibliographic data anchored to ORCID 0000-0002-5419-9292.
- Output, impact, and activity axes are unscored due to insufficient data; the frequent-coauthor list also repeats the researcher's own name and one collaborator, suggesting minor data noise.